Attention-augmented TCN-BiGRU for phreatic line prediction in tailings dams

Abstract The evolution of the phreatic line in a tailings dam is a critical indicator for assessing the structure’s stability. To address the challenges of complex, noisy, and environmentally susceptible monitoring data in real-world engineering, and to achieve rapid, accurate, and synchronous dynamic prediction across multiple monitoring points, this paper proposes a multi-point prediction model. The proposed model, named TCN-BiGRU-MHA, integrates a Temporal Convolutional Network (TCN) and a Bidirectional Gated Recurrent Unit (BiGRU) with a Multi-Head Attention (MHA) mechanism to forecast the phreatic line at four distinct locations within a tailings dam. The model first employs the Maximal Information Coefficient (MIC) for feature selection and processes the dataset using a sliding window approach. Subsequently, the TCN module captures local temporal features, while the BiGRU leverages bidirectional context to respond to abrupt trends. Finally, the MHA mechanism filters and weights key information, enhancing the model’s sensitivity to sudden events and data non-stationarity. The results demonstrate that for the task of phreatic line prediction, the TCN-BiGRU-MHA model achieves a mean Mean Absolute Percentage Error (MAPE) of 0.076% and a mean Coefficient of Determination (R 2 ) of 0.9277. Compared with baseline models such as CNN-LSTM and TCN-BiGRU, the proposed model exhibits superior stability and robustness in the multi-point synchronous prediction task, underscoring its reliability and practical value for phreatic line forecasting.